WEEK OF SEP 20 – 26, 2025

NVIDIA wrote OpenAI a $100B marker the week a 235B vision model shipped with no meter

NVIDIA underwrites OpenAI's next 10 gigawatts, Qwen hands out a 235-billion-parameter vision flagship under Apache 2.0 while keeping its trillion-parameter model closed, a judge signs off on Anthropic's $1.5B book-piracy settlement, and GitHub retires its own extension format for MCP.

The week's largest number was a letter of intent: NVIDIA committed up to $100 billion to OpenAI on Monday as OpenAI contracts at least 10 gigawatts of NVIDIA systems, roughly 4 to 5 million GPUs by Jensen Huang's own count, first phase due on Vera Rubin in the second half of 2026. The day after, OpenAI, Oracle and SoftBank named five more Stargate sites, pushing the program to nearly 7 gigawatts and $400 billion-plus over three years.

The open-weight side of the week moved just as hard. Qwen published a 235-billion-parameter vision-language flagship under Apache 2.0 on Tuesday, native 256K context, queries in 32 OCR languages, weights on Hugging Face at roughly 471 GB. On Wednesday the same team kept its new trillion-parameter Qwen3-Max API-only, no weights, at its Apsara conference — the strong license goes to the model you can download, the meter goes to the one at the top of the leaderboard.

Two legal and platform threads advanced. A federal judge preliminarily approved Anthropic's $1.5 billion settlement with authors on Thursday, the largest publicly reported copyright recovery on record, roughly $3,000 per pirated book. And GitHub began retiring its own Copilot Extensions in favor of MCP servers, trading a proprietary integration format for Anthropic's emergent protocol.

THU · Sep 25, 2025copyrightlitigationtraining dataanthropiccompliance

Judge preliminarily approves Anthropic's $1.5B authors settlement

Judge William Alsup of the Northern District of California granted preliminary approval on Thursday to Anthropic's $1.5 billion class settlement with authors Andrea Bartz, Charles Graeber and Kirk Wallace Johnson — roughly $3,000 per work plus interest across an estimated 500,000 books, with the allegedly pirated datasets to be destroyed. It is the largest publicly reported copyright recovery to date and settles legacy claims only; it does not cover works created or ingested later. Final approval follows the notice and claims process.

The settlement traces directly to Alsup's June ruling, which split the question: training on lawfully acquired books was fair use, but Anthropic's acquisition of millions of titles through Library Genesis and the Pirate Library Mirror was not. The damages land on the acquisition method, not the training itself.

Judge preliminarily approves Anthropic's $1.5B authors settlement
image.cnbcfm.com

WHY IT MATTERS

Provenance is now a priced risk. Keep a per-source ledger for every corpus that touches training, fine-tuning or a retrieval index, and get indemnities in writing from dataset vendors and user-upload terms, because the exposure here was created by how the data was obtained rather than what it was used for. The per-work unit price also gives you a defensible number to put against unlicensed scraping in any build-versus-buy conversation.

WED · Sep 24, 2025mcpcopilot extensionsagent toolingprotocolssupply chain

GitHub retires its Copilot Extensions in favor of MCP servers

GitHub blocked creation of new server-side Copilot Extensions (the GitHub App based flavor) at 8:00 AM PST on Wednesday and will disable all existing ones on 10 November 2025 after a brownout window from 3-7 November. The replacement is Model Context Protocol servers, steered through the GitHub MCP Registry, with per-server allowlisting for IDEs arriving from 28 October. Client-side VS Code Copilot Extensions and ordinary GitHub Apps are untouched.

The stated rationale is the part that matters: one MCP server works with GitHub Copilot, Claude Code and any other MCP-compatible host, removing the build-once-per-platform tax that proprietary extension APIs imposed. GitHub is a major client vendor trading its own integration format for Anthropic's emergent protocol — the clearest signal yet that MCP is becoming the default tool interface for coding agents rather than an Anthropic-flavored option.

GitHub retires its Copilot Extensions in favor of MCP servers
GitHub

WHY IT MATTERS

If you maintain a Copilot Extension you have roughly six weeks to reach functional parity as an MCP server or your integration dies on 10 November, and the upside is portability across every MCP host instead of one marketplace. Treat the registry and per-server allowlists as supply-chain control surfaces: MCP servers run with your credentials and filesystem access, so pin versions, scope each server's tokens to the narrowest capability, and do not let an unpinned third-party server inherit your developer identity.

WED · Sep 24, 2025qwen3-maxclosed weightsagenticalibabapricing model

One day after opening a 235B model, Qwen keeps its trillion-parameter flagship closed

Alibaba launched Qwen3-Max at its Apsara conference in Hangzhou: over 1 trillion parameters, trained on roughly 36 trillion tokens, and available only through the Alibaba Cloud API with no weights published — the first Qwen to cross a trillion parameters and the largest to stay proprietary. Alibaba reports 69.6 on SWE-bench Verified for the Instruct variant and says the Thinking variant, still training at launch, reaches 100% on AIME 2025 and HMMT once tool use and scaled test-time compute are added. It also advertises strong Tau2-Bench results for multi-turn tool use.

The same keynote open-sourced Qwen3-Omni-30B-A3B on Hugging Face and ModelScope, previewed four Wan 2.5 generation models, and upgraded Qwen3-Coder and Qwen3-Image-Edit. Alibaba Cloud CEO Eddie Wu repeated a RMB 380 billion three-year plan for AI and cloud infrastructure and said the company has open-sourced over 300 models since 2023. Read together with Tuesday's release, the split is deliberate: permissive weights at 235B for vision, a closed API at 1T for the text frontier.

WHY IT MATTERS

The strongest weights get the license and the leaderboard leader gets the meter, so the cheapest credible route to frontier-class capability remains an API call to a Chinese provider — and the only real engineering question left is data residency, not capability. Note also that Alibaba is advertising agent tool-use benchmarks rather than chat scores, which is the axis your harness should be evaluated on.

TUE · Sep 23, 2025open weightsqwen3-vlvision languageapache 2.0alibaba

Qwen open-weights a 235B vision-language flagship under Apache 2.0

Qwen published Qwen3-VL-235B-A22B in Instruct and Thinking variants: a mixture-of-experts model with 235 billion total parameters and roughly 22 billion active per token, released under Apache 2.0 on Hugging Face with weights weighing about 471 GB. Native context is 256K tokens, expandable to 1M, with second-level timestamp grounding for long video, 3D grounding for spatial reasoning, OCR across 32 languages, and GUI action planning for computer and mobile interfaces. Qwen reports DocVQA 96.5 in Thinking mode plus MathVista 85.8 and MathVision 74.6, and claims the Instruct build matches or beats Gemini 2.5 Pro on major visual perception benchmarks.

The same day, Qwen shipped Qwen3Guard under Apache 2.0 at 8B, 4B and 0.6B parameters, an openly licensed moderation series in the Llama Guard mould. The license carries no usage cap, no field-of-use restriction and no revenue threshold, so commercial fine-tuning and redistribution are permitted outright. Smaller dense and MoE Qwen3-VL sizes followed in October, which matters because the 235B flagship is not something most teams can hold in VRAM.

WHY IT MATTERS

Document extraction, screenshot-to-UI automation and video indexing can move in-house at a fixed hardware cost, with no per-token meter and no data leaving your network, and the permissive license means you can fine-tune and ship the artefact commercially. Design the vision path as a swappable checkpoint behind your own interface, so today's hosted call and next quarter's self-hosted 8B or 30B build are a config change rather than a rewrite.

MON · Sep 22, 2025infrastructurecomputenvidiaopenaistargate

NVIDIA commits up to $100B to OpenAI for 10 gigawatts of Vera Rubin

NVIDIA and OpenAI announced a letter of intent on Monday for at least 10 gigawatts of NVIDIA systems across OpenAI's data centers, with NVIDIA intending to invest up to $100 billion progressively as each gigawatt is deployed. Huang framed the scale as roughly 4 to 5 million GPUs; NVIDIA's own August earnings maths prices one gigawatt of capacity at $50-60 billion, about $35 billion of it NVIDIA hardware. First phase is targeted for the second half of 2026 on the Vera Rubin platform.

The structure is circular: OpenAI pays NVIDIA cash for chips, NVIDIA buys non-controlling equity at then-current valuations, and the models are trained and served on the same vendor's racks. On Tuesday OpenAI, Oracle and SoftBank added five Stargate sites — Lordstown, Ohio; Milam and Shackelford counties in Texas; Doña Ana County, New Mexico; and an unnamed Midwest site — bringing Stargate to nearly 7 gigawatts of planned capacity and more than $400 billion of investment over three years. OpenAI's announcement pegs its user base at over 700 million weekly active users.

WHY IT MATTERS

Frontier compute is now contracted years ahead and financed on the supplier's balance sheet, which makes your inference ceiling a function of deals you are not party to. Practically: expect the cheap tier to keep thinning from the top, and treat multi-provider routing and a self-hosted fallback as architecture rather than a rainy-day plan, because capacity crunches will surface as rate limits and price floors before they surface as outages.